predict: move bench service to deepbook-services

[pei-mysten]

Summary

• Removes crates/bench/ and docker/predict-bench/ — the bench service moves to deepbook-services alongside the other deepbook-* services
• Removes crates/bench from the workspace Cargo.toml; Cargo.lock regenerated without bench's transitive deps
• Keeps docker/predict-sim/ and packages/predict/simulations/ here — the sim container's entrypoint clones packages/predict/simulations/ at the requested SHA per benchmark run, so they are tightly coupled to the protocol code by design

Key decisions

• Why split bench from sim: bench is pure ops infra (Axum service + k8s Job orchestration), sim is protocol-coupled (clones the source tree). Different release cadences, different reviewers, different concerns
• predict-bench.yml workflow stays here — it only references the deployed bench service by URL, no local paths
• Cargo.lock changes are mechanical (removing bench's transitive deps from the lock)
• sim.ts is unchanged in this PR. The loadScenario(process.env.SCENARIO_PATH) change needed to let the sim pick up the GCS-downloaded scenario CSV is being handled in a separate PR by another teammate who owns that area. Until that lands, benchmarks against this PR's commits will silently ignore the bucket CSV and use the in-repo scenario_mar6_1000mints.csv, which matches the pre-bench-move behavior — so this PR is fully backwards-compatible and safe to land independently

Coordination

Part of a coordinated move across three repos:

• deepbook-services#7: adds the bench code
• sui-operations#8080: switches Pulumi to build the bench image from deepbook-services (sim image still builds from here) + provisions GCS scenario bucket
• this PR: removes the bench code here

Recommended landing order: deepbook-services first → sui-operations second → this PR last (avoids any broken-deploy window — the bench code exists in both repos briefly, sui-operations switches sources, then this PR cleans up). The companion sim.ts change is decoupled and will land on its own schedule.

Test plan

• cargo check clean on the workspace after removal
• No behavior change for existing benchmarks (sim.ts unchanged → falls through to the in-repo scenario CSV as before)
• After all three PRs land, manual benchmark via the deployed service still succeeds end-to-end against the in-repo scenario

🤖 Generated with Claude Code